Test Problem

This document presents a description of the test problem and the corresponding data set
used to compare ensemble-based methods in the paper: “Investigation on the sampling
performance of ensemble-based methods” by Alexandre Emerick and Albert Reynolds.

The objective of this data set is to allow other research groups to reproduce the results in
the original paper and to test their own implementation and methods. Besides the true
model, the data set includes the results obtained with Markov chain Monte Carlo (MCMC)
using a very long Markov chain (20 million proposals). The results from this Markov chain
were used as reference distributions to compare the results of the ensemble-based methods.

Description

The test problem is a one-dimensional reservoir model under waterflooding (Fig. 1). The
number of gridblocks is 31 and the dimensions of all gridblocks are 50 ft × 50 ft × 50 ft.
The model parameters are gridblock log-permeabilities, ln(k). The “true” permeability field
(Fig. 2a) was generated using an exponential covariance function with a practical range
corresponding to the size of 10 gridblocks. The prior mean of ln(k) is 5.0 and prior variance
is 1.0 for all gridblocks. The porosity is constant and equal to 0.25; the oil viscosity is 2 cp
and water viscosity 1 cp. The initial reservoir pressure is 3,500 psi and the compressibility
of oil, water and rock are 10–5 psi–1, 10–6 psi–1 and 5×10–6 psi–1 respectively. In this
synthetic reservoir, there is a water injection well in the first gridblock which is operated at
a constant bottomhole pressure of 4,000 psi. In the last gridblock, there is a producing well
operated at a constant bottomhole pressure of 3,000 psi. The observations correspond to
gridblock pressures at a monitor well located in the center of the reservoir (gridblock
number 16). The historical period corresponds to 360 days, with one pressure measurement
every 30 days, which results in 12 data points. We added random Gaussian noise with a
standard deviation of 1 psi to the data predicted by the true model to define the
“measurements.” We chose a test case with few data points and small measurement errors
to make the problem more challenging for data assimilation. The historical period was
defined such that we have water breakthrough at the monitor well but not at the producing
well. Fig. 2b shows the water saturation distribution at the end of the history (360 days) and
at the end of a forecast period (750 days).

Fig. 1: Gridblocks and well locations.

10000 0.7

0.6
Permeability (mD)

1000 0.5
Sw

0.4

100 0.3

0.2

10 0.1
1 6 11 16 21 26 31 1 6 11 16 21 26 31
Gridblock Gridblock

(a) (b)
Fig. 2: (a) Permeability of the true model. (b) Water saturation at the end of the historical period (in blue) and
at the end of the forecast period (in red). The vertical dashed line indicates the position of the monitor well.
Data set
Description File name Comments

Problem description Problem_description.pdf This file

Gridblock permeability in mD of
True permeability PERMI_TRUE.inc
the true model.
File with prior covariance of log-
Prior covariance matrix Cm.txt
permeability (ln k).
Set of files to run reservoir
simulations. It includes the
following files
- model.dat
Simulation files model.zip - null.inc
- PERMI_TRUE.inc
- equilibrium.inc
- initialization.inc
- recurrent_data.inc
- observed_data_press.obs
2D-2 phase reservoir simulator
used in the test problem. It
Reservoir simulator TwoFlow.exe
includes adjoint-based gradient for
history matching.
Excel file with main result of
MCMC, which are the reference
sampling results. This file includes
the distribution of permeability
MCMC results MCMC_results.xls (mean, standard deviation and
percentiles P2, P25, P50, P75 and
P98) and the distributions of
predicted data (mean, s.dev and
percentiles.)
10 initial ensembles of
Prior ensembles ens_xx.zip permeability used in the original
paper.
Observed data observed_data.txt Observed data with noise.

November 27, 2011

Alexandre Emerick (aemerick@gmail.com)
Albert Reynolds (reynold@utulsa.edu)